Repository navigation
Interpretable Learning
The working group on Interpretable Learning explored methods to generate predictive models that can also generate an explanation.
There are several machine learning methods that utilize Linked Data, or knowledge graphs, to generate predictions of biologically meaningful relations, provide similarity-based operations such as approximate searches or clustering, or generate features for other machine learning models. These methods include knowledge graph embeddings (), graph convolutional neural networks (), and others (). The majority of these methods utilizes neural network models or other vector space operations and are not easily amenable to generating explanations of why a prediction has (not) been made.
There are several machine learning methods that result in interpretable models. We consider a model to be interpretable if it provides a symbolic description that is directly supporting the predictions. Logic programs are the most common form of such symbolic descriptions. Consequently, we tested the application of inductive logic programming and probabilistic knowledge representation and reasoning methods.
We explored the use of Markov Logic Networks for predictive analysis on knowledge graphs in biology. Markov logic defines a probability distribution over model structures...
Our application was to predict protein-protein interactions where the test dataset came from String database. There are quite large memory requirements for these tools, so data size had to be limited for now. We used E coli as our target because there is a relatively low number of proteins and interactions for E coli in String.
These methods worked quite well and predict interactions, with a likelihood.
Todo: Need to explore some of the different sampling and prediction methods.
We also explored the use of logic with decision trees for prediction of relationships on biological knowledge bases. Decision Tree techniques provide a logic progam ( a decision tree ) and explanations as rules for every prediction.
We applied TILDE, a first-order learning approach, to predict PPI for yeast. The calculation was memory demanding and it took more than 24h of computation for an organism with a total of 6,275 genes and that at least possesses 23% homologous genes to humans Liu et al.. We obtained predictions supported by explanations defined by a decision tree. We plan to compare the performance of these tools with neural network techniques.
Todo: Need to investigate and optimize algorithm performance and implementation for biomedical big data and specifically for human databases.
https://github.com/bio-ontology-research-group/interpretable-learning/